CRoS improves CoT causality with scrambled reasoning steps
Original title: CRoS: Causality Reward on Scrambled CoTs Improves Accuracy and Reasoning Faithfulness
CRoS perturbs CoT order without altering content to measure causality, achieving 46.8% accuracy on unseen datasets with Gemma-3-4B.
CRoS provides a CoT causality metric that scrambles reasoning step order to avoid information changes from perturbations. On Reasoning Gym with Gemma-3-4B it outperforms baselines in out-of-domain accuracy, CoT quality, and faithfulness. Hint acknowledgment reaches 34.0% versus 20.2% for existing metrics.
- 46.8% accuracy on five unseen logic datasets
- 43.4% vs 41.4% CoT-Pass@K average
- 34.0% vs 20.2% hint acknowledgment score
- preserves information within each CoT step
- AI models
- Gemma-3-4B
From the original article
1 min read at arxiv.orgReinforcement learning with verifiable rewards (RLVR) has become the standard method for training language models to reason. However, RLVR typically relies solely on outcome correctness, which does not guarantee that the CoT causally contributes to the final answer. Without this connection, outcome rewards may fail to reinforce high-quality reasoning.
By Xiao Yu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.